Suggestions
Parker Odrich
Investor at Point72 Ventures
Parker Odrich is the Vice President at Communitas Capital, a position he assumed after a successful career in investment and consulting. He holds a Bachelor of Science in Management Science & Engineering from Stanford University, where he graduated with high honors.
His professional background includes significant roles in investment and analysis. Prior to joining Communitas Capital, Parker was an Associate at Alvarez & Marsal Capital, where he worked on strategic investments and led the firm's AI Task Force, evaluating AI-related opportunities. He also served as an Analyst in Private Equity Services at Alvarez & Marsal, focusing on operational due diligence and corporate transformation across various industries.1
Parker has experience as an Associate Investment Analyst at JetBlue Technology Ventures and has held internships at Sweep and Bertram Capital, where he contributed to product development and research initiatives. His expertise spans management, finance, technology, and data consulting, making him a well-rounded professional in the investment sector.1
Highlights
I view alignment, the process of encoding human values and goals into LLMs to make them reliable, as a critical concept in developing chat-based LLM applications, especially in matching responses with expectations for specific use cases. A trend I’ve been noticing in building chat-focused LLM applications is an alignment technique known as Direct Preference Optimization, where the LLMs are optimized directly. This is evidenced by Mistral training its Mixtral 8x7B Instruct model using DPO instead of RLHF (https://lnkd.in/gjdfpKdv). So why do I think this is important? In my view, large scale implementations of LLMs benefit more from DPO as you're directly optimizing the model opposed to RLHF, which introduces the reinforcement learning component; RLHF tuning inherently increases complexity and is more computationally heavy. I believe it is important now as large corporates are figuring out their "AI Strategy" and need to be able to monitor their LLMs for ROI tracking efficiently. I view this adoption as promising in making model development more cost effective, especially with the plethora of open source training data and synthetic data, and monitoring less complex & burdensome post-implementation. Feel free to ping me at [email protected] with any thoughts or paper recommendations around this exciting topic!
I’m happy to share that I’m starting a new position as an Investor at Point72 Ventures!
